Skip to content
/ tidbytes Public

Memory manipulation reimagined with bit addressing

License

Notifications You must be signed in to change notification settings

Pebaz/tidbytes

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Tidbytes Bit Manipulation Library

Memory manipulation reimagined with bit addressing

Referring to the ninth bit of a region of memory should be trivial. It is, isn’t it? It turns out that this is not as straightforward as it might seem. Depending on the byte order, (referred to under the canonical moniker “endianness”), the ninth bit may appear to the left or the right (not accounting for mixed endianness) of the first bit (bit zero). The first byte should be the rightmost byte for numeric data, but not when reading from a file for example. In this case the first byte would be the leftmost byte. Because of all of this, I designed Tidbytes to work specifically with bits as the primary addressable unit of memory to provide clarity when working with memory from diverse sources.

The purpose of Tidbytes is to allow bits to be placed precisely where they are wanted. In the pursuit of mapping this ideal to idiomatic types, some in-built concepts were rediscovered. There really seems to be some fundamental types in relation to mapping numeric data to bits. "Type" here refers to a collection of related operations that either assumes metadata about an input or requires metadata as a meta input. This represents an orientation that points away from the operation and towards an operation respectively. Some concepts that relate to operations that I've rediscovered are:

  • Unsized data (inputs can be mapped to unlimited output data)
  • Sized data (inputs can be mapped to output data limited by region size)
  • Natural data (raw/untyped/uninterpreted/unmapped memory)
  • Numeric data (mathematical identity or quantity)
  • Unsigned numbers (axis with one polarity)
  • Signed numbers (axis with two polarities)

These types are built into modern processor architectures via Bytes & Indices. Tidbytes takes these natural computing concepts and provides 2 APIs:

  • A lower level API designed to be easily re-implemented in other languages
  • A higher level API designed to be as seamless as possible for Python programmers
graph TD;
    subgraph Operation API Design
        direction LR
        mem1{{Bytes}}
        mem2{{Bytes}}
        index{{Index}}
        mem1 -- input --> op -- output --> mem2
        index -- input --> op
    end
Loading

Each operation in the lower level API takes Bytes & Indices as input and produces Bytes as output or Indices if it is a meta operation. Using just Bytes and Indices, all memory manipulation operations can be implemented because Indices can represent offsets, lengths, ranges, slices, and (un)signed numbers.

graph RL;
    subgraph API Design Layers
        direction LR

        idiomatic("
            Language Specific Idiomatic Interface
            (High Level API)
        ")

        natural("
            Language Agnostic Natural Interface
            (Low Level API)
        ")
    end

    natural_type{{"
        Natural Memory Type
        (Language Agnostic)
    "}}

    idio_type{{"
        Idiomatic Memory Type
        (Language Specific)
    "}}

    idiomatic ==> natural
    natural_type --> natural
    idio_type --> idiomatic
    natural_type --> idiomatic
Loading

Idiomatic Higher Level API

The higher level "idiomatic" API provides these main types: Mem, Unsigned, and Signed. The Mem type is the base type for all the other types and provides the majority of the higher level functionality that will be comfortable to Python programmers. It upholds contracts on invariants that are unique to unsized or sized natural data in the form of raw bytes. The Unsigned type upholds contracts on invariants unique to unsized or sized unsigned numeric data. The Signed type upholds contracts on invariants unique to unsized or sized signed numeric data.

from tidbytes.idiomatic import *

mem = Mem[8]()  # <Mem [00000000]>
mem[0] = 1      # <Mem [10000000]>
mem[0:3]        # <Mem [100]>
mem += mem      # <Mem [10000000 10000000]>
mem[1::8]       # <Mem [10000000]>  Step by bytes

U2 = Unsigned[2]  # Type alias
def add(a: U2, b: U2) -> U2:
    return U2(a) + U2(b)

# Underflow and Overflow checks:
add(U2(2), U2(2))  # Error: 4 doesn't fit into bit length of 2 (min 0, max 3)

num = add(U2(2), U2(1))  # <Unsigned [11]>

# Fundamental type conversion support
int(num)  # 3
str(num)  # '11'
bool(num)  # True
float(num)  # 3.0

Each of the higher level types can be constructed from most of the types shown here, although some types like Unsigned cannot be given a negative big integer or a Signed memory region.

%%{ init: { 'flowchart': { 'curve': 'basis' } } }%%
graph LR;
    subgraph "Many Input Types, Only One Output Type"
        subgraph Higher Level Types
            _6{{"list[bit]"}}
            _7{{"list[byte]"}}
            _8{{"list[u8]"}}
            _9{{"bool"}}
            _10{{int}}
            _4{{ascii}}
            _5{{utf8}}
        end

        %% Order matters here, they are interlieved
        _1{{u8}} --- Mem
        _10{{int}} ---- Mem
        _3{{f32}} --- Mem
        _5{{utf8}} ---- Mem
        _2{{i64}} --- Mem
        _6{{"list[bit]"}} ---- Mem
        _11{{u16}} --- Mem
        _7{{"list[byte]"}} ---- Mem
        _12{{u32}} --- Mem
        _8{{"list[u8]"}} ---- Mem
        _13{{u64}} --- Mem
        _9{{"bool"}} ---- Mem
        _14{{i8}} --- Mem
        _4{{ascii}} ---- Mem
        _15{{i16}} --- Mem
    end
Loading

Natural Lower Level API

The lower level "natural" API is much more verbose and not designed to follow Python idioms but rather to conceptually model C-equivalent semantics when possible. Although demonstrated here, the lower level API is not designed with any user experience features by design, as that is the express role of the higher level API. The goal of the lower level API is to make it easier to port to new languages.

Another feature that the natural API provides to the higher level API is the data transformation functions for converting to and from bytes. These are called codecs as they either encode or decode types or memory. These codecs are used by the higher level API for all supported idiomatic types such as bool or int.

from tidbytes.natural import *
from tidbytes import codec

# The underlying backing store for bits is a list of list of bits.
# Obviously this is not the highest performance choice but the logical clarity
# it provided for this reference implementation could not be ignored
mem = codec.from_bit_list([1, 0, 1], 3)
assert mem.bytes == [[1, 0, 1, None, None, None, None, None]]

mem = op_concatenate(mem, mem)
assert mem.bytes == [[1, 0, 1, 1, 0, 1, None, None]]

mem = op_truncate(mem, 3)
assert mem.bytes == [[1, 0, 1, None, None, None, None, None]]

assert meta_op_bit_length(mem) == 3

# This codec assumes signed memory as input since Python ints are always signed
assert codec.into_numeric_big_integer(mem) == -3

Lower level API operations:

  • op_identity: maps a memory region to itself (multiplies by 1)

  • op_reverse_bytes: transforms little endian to big endian and vice versa

  • op_reverse_bits: reverses the bits of each byte while maintaining byte order

  • op_reverse: reverses both bits and bytes, effectively flipping the entire region

  • op_get_bits: slices out a range of bits into another range of bits

  • op_set_bits: sets a range of bits with another range of bits

  • op_truncate: remove additional space if greater than provided length

  • op_extend: fill additional space with a value if less than provided length

  • op_ensure_bit_length: fill or remove space if less or greater than length

  • op_concatenate: combine two memory regions from left to right

Lower level API codecs:

The difference between numeric and natural is that the output for numeric codecs preserve semantic bit order. The byte 00000001 may be the numeric quantity 1 or the raw memory region 10000000. Natural stores the byte by semantic order and numeric stores from right to left bit order which is more appropriate for integers.

  • from_natural_u8: convert a non-numeric raw byte into memory
  • from_natural_u16: convert a non-numeric 2-byte region into memory
  • from_natural_u32: convert a non-numeric 4-byte region into memory
  • from_natural_u64: convert a non-numeric 8-byte region into memory
  • from_numeric_u8: convert a 1-byte unsigned number into memory
  • from_numeric_u16: convert a 2-byte unsigned number into memory
  • from_numeric_u32: convert a 4-byte unsigned number into memory
  • from_numeric_u64: convert an 8-byte unsigned number into memory
  • from_natural_i8: convert a raw 1-byte signed number into memory
  • from_natural_i16: convert a raw 2-byte signed number into memory
  • from_natural_i32: convert a raw 4-byte signed number into memory
  • from_natural_i64: convert an raw 8-byte signed number into memory
  • from_numeric_i8: convert a 1-byte signed number into memory
  • from_numeric_i16: convert a 2-byte signed number into memory
  • from_numeric_i32: convert a 4-byte signed number into memory
  • from_numeric_i64: convert an 8-byte signed number into memory
  • from_natural_f32: convert a raw IEEE754 single precision float into memory
  • from_natural_f64: convert a raw IEEE754 double precision float into memory
  • from_numeric_f32: convert an IEEE754 single precision float into memory
  • from_numeric_f64: convert an IEEE754 double precision float into memory
  • from_natural_big_integer_signed: convert into memory
  • from_natural_big_integer_unsigned: convert into memory
  • from_numeric_big_integer_signed: convert into memory
  • from_numeric_big_integer_unsigned: convert into memory
  • from_natural_float: convert a raw IEEE754 float into memory
  • from_numeric_float: convert an IEEE754 float into memory
  • from_bool: convert a boolean into memory
  • from_bit_list: convert an int list of 0 or 1 into memory
  • from_grouped_bits: convert a list of lists of 0 or 1 bits into memory
  • from_bytes: convert a list of ints in range 0-255 into memory
  • into_numeric_big_integer: convert numeric memory into idiomatic integers
  • into_natural_big_integer: convert natural memory into idiomatic integers

Here is a diagram that shows dependencies between operations so that they can be re-implemented in order in another language:

mindmap
    root("
        Tree Of Re-
        Implementation
    ")
        op_transform
            op_identity
            op_reverse
            op_reverse_bytes
            op_reverse_bits
        op_get_bits
            op_get_bit
            op_get_byte
            op_get_bytes
        op_set_bits
            op_set_bit
            op_set_byte
            op_set_bytes
        meta_op_bit_length
            meta_op_byte_length
        op_ensure_bit_length
            op_truncate
            op_extend
        op_concatenate
Loading

The Ninth Bit

↪ Reasoning about the ninth bit within the context of programming computers is not as straightforward as it might seem. It entails preconceived notions on the part of the programmer about how the runtime CPU architecture loads bits into registers as well as assumptions around the origin of those bits. The ability to refer to singular bits is not a capability natural to modern computer architectures due to byte addressing, but there is utility in doing so nonetheless. Some applications of referring to bits is data format encoding, structure bit fields and layout, and machine code instruction encoding. As such, to get around the limitation of bytes as the lone addressable unit of memory, bit locations are calculated at runtime through the use of arithmetic and bit shifting. This is due to the limitations in the available instructions to the assembly programmer. However, as will be seen below, thinking past this limitation in a higher level of software provides logical coherency that could aid application programmers when integrating with lower level libraries, operating systems, and hardware. Although this reference library implementation was primarily designed for correctness and ergonomics rather than performance, it may be useful for prototyping, testing, or other limited bit manipulation tasks.

Glossary of Terms

  • Natural: Refers to the bit (and most commonly) byte order of a given processor architecture: the memory universe. It is from the point of view of the host.

  • Foreign: Refers to a memory universe of another processor host, regardless if it exactly matches.

  • Origin: Refers to which memory universe a memory region was allocated.

Identity Order

Tidbytes is based off of the concept of Identity Order which means the first bit is always the leftmost bit of the leftmost byte.

When both the bit and byte order of a region of memory is left to right, this is called “Identity Order” in Tidbytes. The following bits are in identity order. Performing the corresponding memory transformation on them (left to right bit and byte transform: op_identity) will do nothing because they are already in left to right bit and byte order.

------->
00000000 00000000
->

When using the memory transformation operations op_identity, op_reverse, op_reverse_bits, op_reverse_bytes, the result is always identity bit and byte order. However, to vizualize this, different notation is used:

                ------->
Identity        10000000 01000000
                ->

                <-------
Reversed        00000010 00000001
                <-

                ------->
Reversed Bits   00000001 00000010
                <-

                <-------
Reversed Bytes  01000000 10000000
                ->

Amazingly, performing any of those 4 operations transforms the shown memory regions into identity bit and byte order. This is the most useful memory order for bit reading (indexing, offsetting) and writing (set, concat, extend, truncate) operations because it matches most mathematical notation such as equations and cardinal graphs. For memory in identity order, it is unlikely to be semantically meaningful in a primitive (scalar) way. Generally, memory in identity order tends to be for everything except for directly storing data.

For identity order memory, the first and ninth bits are the leftmost bit of the leftmost byte and the leftmost bit of the second byte from the left. Numeric data does not index the same way as raw memory such as bit-fields. Bits go from right to left and bytes go either left to right or right to left based on memory universe byte order. Being able to index from the correct direction allows smaller numbers to be sliced out of large fields such as a 3-bit integer sliced from an 8-bit byte.

Identity Order:

Byte: 1st      2nd
       |        |
       -------> |
       V        v
       00000000 00000000
       ^        ^
       ->       |
       |        |
Bit:  1st      9th
Reversed Bits (Little Endian):

Byte: 1st      2nd
       |        |
       -------> |
       V        v
       00000000 00000000
       <-     ^        ^
              |        |
              |        |
Bit:         1st      9th
Reversed Bytes:

Byte: 2nd      1st
       |        |
       <------- |
       V        v
       00000000 00000000
       ^        ^
       ->       |
       |        |
Bit:  9th      1st
Reversed (Big Endian):

Byte: 2nd      1st
       |        |
       <------- |
       V        v
       00000000 00000000
       <-     ^        ^
              |        |
              |        |
Bit:         9th      1st

The difficulty in sorting all this out by hand per origin and destination bit and byte order is the primary motivation for the creation of Tidbytes.

Memory Origin And Universes

The most common byte orders are left-to-right (little endian) and right-to-left (big endian). Big endian matches how numbers are written while little endian allows zero-indexing of bytes.

Although it is commonly thought that bits always go from right-to-left, on occasion they also go from left to right. This can be the case when slicing out bit fields that are smaller that a byte from structs or that otherwise cross byte boundaries. In such cases, is the first bit on the far right or on the far left? This is an intriguing duality. Memory origin does seem to matter then. When considering the entire struct, the first bit is always the leftmost bit of the leftmost byte. When considering numeric data, the first bit is the rightmost bit of the rightmost byte. Strange.

The use case for transforming between memory universes often comes up when reading from a file or a network socket. When reading bytes from a file, they are read (logically) from left to right one at a time. These bytes come from an entirely separate memory universe: the universe of the file format. Once they have been read into memory, they are now within the memory universe of the program, although they have not yet been transformed to identity order.

Amazingly, simply applying a foreign memory region’s bit and byte order as a transformation on itself produces that same region in identity order, easily usable by the host program. This is a surprising insight which ensures that the “first bit” is always the leftmost bit of the leftmost byte.

graph TD;
    subgraph Memory Universe Boundaries Require Transformation Operations
        direction LR

        _1(((Memory Universe 1)))
        _2(((Memory Universe 2)))
        op_transform
        _1 <===> op_transform <===> _2

        subgraph op_transform
            op_identity
            op_reverse
            op_reverse_bytes
            op_reverse_bits
        end
    end

    style op_transform fill:#9893bf,stroke:#5c5975
Loading

Here's an example of typical bit and byte orders for various bit manipulation tasks:

Read From Order
Bytes File Left To Right
Bytes Struct Left To Right
Bytes Number Left To Right (little endian)
Bytes Number Right To Left (big endian)
Bytes Network Right To Left (multi-byte field)
Bits File Left To Right
Bits Struct Left To Right
Bits Number Right To Left

TLDR;

Take the origin of a region of memory and perform the corresponding memory transformation operation to map it to identity order in the natural host CPU memory so that bits and bytes are always from left to right.

Foreign Order Natural Order Transformation Operation
Big Endian Little Endian op_reverse_bytes
Little Endian Big Endian op_reverse_bytes
Big Endian Big Endian op_identity (no op)
Little Endian Little Endian op_identity (no op)

So What Truly Is The “Ninth Bit”

By taking a foreign memory region and applying it’s own bit and byte order as a transformation upon itself it yields a region with identity memory order, wherein the “ninth bit” is always the leftmost bit of the second byte from the left.

Desirable Future Additions

  • Make the indexed_meta type work with Mypy type checking
  • Implement more operations: shift, shift with rotate, all fundamental Python bitwise operations. Xor, any(), all(), logical.
  • Choose codec by keyword: Mem(string='asdf'), Mem(hex='0xFF', len=100)
  • Rewrite the natural memory type to use bytes instead lists of lists of bits
  • Describe exact bit layout for a data type (struct)
    class Struct(Mem):
        def __init__(self, *args, **kwargs):
            self.buffer = Mem(sum(args))
    
    class Union(Mem):
        "Works similarly"
    
    class Ieee754Single(Struct):
        Sign: Unsigned[1] = 0
        Exponent: Unsigned[23] = '0101010'
        Mantissa: Unsigned[12]
    
        def __float__(self):
            return float(self.buffer)
  • Implement Union type
  • Type layout with default values (foo: Signed = 3)
  • Types with templates that can be filled in later (like inst constant in ASM)
  • Effective sizeof & effective alignof (not the same as bit len)
  • Bit-level Cursor API for parsing data structures
  • Cursor API bit-level read ahead and skip for reading small integers like u3
  • File-like bit-level write API for assembling (write(bits=4, value=3))
  • Store and query endianness of region
  • Implement ALU operationss like XOR, etc.
  • Match bit patterns like masks and interleaved bits (better than OR)
  • Convert internal representation to use a byte array, not a list